Video AI Generator

Six prompts, six clip slots, nothing borrowed

Every page selling video generation shows you a still image. That is the strangest habit in this whole category. Below are six slots for this page's own clips, each captioned with the literal prompt that will produce it and the model that will run it. A slot showing a placeholder frame has no file behind it yet and the caption says so rather than pretending. When a clip is generated and uploaded, the slot plays it and the caption swaps the badge for the real time it took to generate.

Text to video AI, prompt by prompt

Pick a prompt and the matching slot loads immediately. No model runs on this page and none ever will: a published run is a file served from this domain, so there is no queue, no credit and no account. Where a run has not been published yet the picker says so instead of playing something borrowed.

Pick a prompt. Nothing generates in the browser. The run behind this prompt has not been published yet, so the slot shows its poster frame until the file lands. Nothing is queued and nothing is charged.

• Drone over a city
• Espresso pour
• Runner at dawn
• Product orbit
• Hands in a kitchen
• Rain on glass

Prompt queued

Aerial drone shot pushing forward over a coastal city at golden hour, long shadows on the water, slow forward dolly, 35mm, shallow haze.

model

Veo 3

frame

16:9

file

/media/home-city-drone.mp4

state

run not published yet

Text to video is the input mode most people mean when they search for this. You write a description, the model produces frames that match it. The other common input is a still photo, which is a different problem with different failure modes, covered on the image to video page.

What a video generator AI does, in one paragraph

A video generator AI predicts frames. You give it a condition, either text or an image, and it produces a short sequence of frames that satisfies that condition and stays consistent with itself from one frame to the next. That last part is the hard bit and it is why clips are short. Consistency decays as the sequence gets longer, so most models stop at four to ten seconds before the subject starts to drift.

Everything else the tools sell sits on top of that core loop. Camera controls constrain the motion. Native audio adds a second output head that emits sound aligned to the picture. Script mode chops your writing into scene prompts and runs the loop once per scene. The loop is the product. The rest is packaging.

What free actually means here

Every tool in this category calls itself free and almost none of them define the word on the page. The unit that matters is seconds of usable, watermark free video per month at zero cost. Here is what each tool publishes on its own page, read in August 2026, with blanks left blank where a page does not say.

Read the second column carefully. Three generations a day at three seconds is nine seconds of video a day. Ten minutes a month of avatar video is not the same product as ten minutes of generated footage. A daily allotment of credits with no published exchange rate is not a number at all. The arithmetic is done properly on the free text to video comparison, and the account and watermark side is compared on the no sign up page.

AI video maker versus video AI generator

These two phrases return almost the same search results and mean different things. A generator makes footage that did not exist. An AI video maker assembles footage that already exists, adds captions, music and transitions, and often calls the assembly step AI as well.

The distinction matters when you are picking a tool. If you have footage and need it cut, a maker is the right product and a generator is a slow, expensive way to get worse results. If you have nothing but an idea, a maker leaves you searching a stock library and a generator gives you the shot. Several products in this market are both, which is why the category reads as confusing.

The models behind the tools

Most of the products in this category are interfaces over the same handful of models. Knowing which model a tool runs tells you more about the output than the marketing on its page.

Picking a model for what you are shooting

The table above says what each family is good at. This is the same information as a decision, which is what you actually need when you are about to spend a take.

• You have a photo and want it to move. Kling. Image to video is where it is strongest and the framing stays stable.
• You need one shot that must feel real, with sound. Veo 3. Native audio is the thing you cannot add afterwards in a minute.
• You need vertical, fast, stylised. Seedance. It handles motion and 9:16 framing better than it handles fine detail, which suits the format.
• You need texture, atmosphere, or a background that is not the subject. Wan. Cheap, reliable, and nothing in the frame has an identity to lose.
• You need physical plausibility over several seconds. Sora 2, if you have access to it. Otherwise shorten the shot and use anything else.
• You are cutting eight clips together. Any silent model, one look, one audio bed built underneath. Mixing native audio models across a cut is the most common way to make a sequence sound wrong.

One habit is worth more than the whole list. Generate the same prompt on two models before you commit to a project. Ten minutes there saves you from finding out at clip eleven that you picked the wrong one.

What this cannot do yet

This section exists because no competing page has one. If you plan a project around video generation without knowing these, you will lose a day finding them out yourself.

• Clip length. Four to ten seconds per generation is the working range. There is no single prompt that returns a two minute video. Long videos are stitched.
• Faces past about four seconds. Identity drifts. Eyes and mouths are the first to go, then the jawline. Cut before it happens rather than fighting it in the prompt.
• Hands. Fingers merge and multiply during fast motion. Frame hands close and slow, or keep them out of shot.
• Text in frame. Signs, labels and screens rewrite themselves mid clip. If your shot needs readable words, add them in post as an overlay.
• Lip sync on generated speech. Models that emit dialogue rarely land the mouth shapes. It reads as a badly dubbed film. Voiceover over cut footage is more reliable.
• Character consistency across shots. The same person prompted twice comes back as two different people. Reference images help and do not solve it.
• Counting and physics. Ask for five birds and you get somewhere between three and eight. Ask for a glass to break exactly on impact and it may break a beat early.
• Editing an existing take. You cannot ask for the same clip with one thing changed. Re-running the prompt gives you a new clip, not a revision.

How the pricing actually works

Nearly every tool prices in credits, and credits are spent per second of output rather than per generation. That means a ten second clip costs roughly twice a five second clip, and a failed take costs the same as a good one. Budget for the takes you throw away, because that is where most of the spend goes.

Two things move the rate: the model and the resolution. A silent image to video model at 720p is the cheapest thing on any of these platforms. A native audio model at 1080p is the most expensive. Published rates we found while researching this page start around two cents per second of output on an API, and the consumer plans work out higher once you account for failed takes.

Your first ten generations

Everyone spends their first session on the same mistakes. Here is the shortcut, in the order the mistakes normally happen.

1. Start with an easy subject

   A landscape, a texture, a product on a plain background. Not a person. You are learning how the tool responds, and a hard subject tells you nothing except that hard subjects are hard.

2. Name the camera move

   If you do not, every clip comes back as a slow push in. Pan, orbit, dolly, tilt, static. One word changes the whole result and most people never type it.

3. Ask for one action

   Two actions in a prompt roughly halves the number of takes you can use. Split them into two clips and cut between them, which is what a real edit does anyway.

4. Say what should not move

   These models animate everything by default. If the background should be still, that is a thing you have to ask for.

5. Generate low, finish high

   Iterate at the cheapest resolution the tool offers and only re run the prompt you like at full quality. Most of your spend otherwise goes on high resolution versions of ideas you rejected.

6. Keep the takes you rejected

   A take that failed as a hero shot often works as a two second cutaway. Deleting them is throwing away footage you already paid for.

Mistakes that cost the most

• Prompting with quality words. Cinematic, 8K, masterpiece, award winning. They do almost nothing. Specifics do everything.
• Generating long and hoping. The last two seconds of a ten second clip are the worst two seconds of it. Generate what you need and cut early.
• Ignoring frame rate. Clips from different tools come back at different rates and stutter at every cut. Normalise on import, not after the edit.
• Expecting revisions. There is no version of this clip but with the door open. Every run is a new clip. Prompt for what you want the first time.

The words this category uses

Most of the confusion around these tools comes from vocabulary rather than from technology. These are the terms that appear on every product page.

How to read a page like this one

There are a lot of pages selling video generation and most of them tell you the same four things. These are the checks worth running on any of them, including this one.

• Does it show video? A page selling video generation that illustrates itself with still images has not tested what it is selling.
• Are the prompts printed? A clip without its prompt is a marketing render. With the prompt it is evidence.
• Is the free tier quantified? Free with no number attached means the number is bad.
• Are the limits published? Every one of these models fails in the same predictable places. A page that lists none of them has not used the product for long.
• Is anything dated? This category changes monthly. An undated comparison is a snapshot of a moment nobody can identify.